This PDF file contains the front matter associated with SPIE Proceedings Volume 12791, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

In combinatorial optimization, Traveling Salesman Problem (TSP) is a well-known issue and one of the key problem in the logistics industry in recent years. The speed, accuracy, and generalization ability of the traditional method and the algorithm design of the specific problem are greatly affected. With the wide application of deep reinforcement learning (DRL) in industry, the automatic design of learning algorithms using DRL models has become a recent research hotspot. This paper proposes a hybrid network model with a spatial attention mechanism to resolve large-scale TSP in order to enhance the generalization capability of DRL-based model on large-scale TSP. Spatial attention is beneficial to capture the global connection between nodes and then to calculate and extract key features by weighting all local features, which is beneficial to improving the generalization ability of the model. The experimental results demonstrate that our model can significantly increase the route problem's optimization effectiveness.

In the detection of mining vehicles and pedestrians, the classical object detection model is not suitable for micro-small hardware equipment due to its large size and large number of parameters.However, the existing lightweight models are difficult to balance detection accuracy and real-time. In this study, a lightweight detection algorithm Depthwise MobileNetv3 -YOLOv4based on improved Mobilenet-YOLO is proposed to implement the compression of the network model. At the same time, the use of deep separable convolution on the detection head improves the ability of the network to pay attention to the target feature information, Finally, the Siou activation function is used to improve the problem of slow network convergence. The experimental results show that the Depthwise MobileNetv3-YOLOv4 network parameters are reduced by 90% compared with the original network under the condition of ensuring accuracy, and compared with other network models, Depthwise MobileNetv3-YOLOv4 is more suitable for vehicle volume.

The 17 Sustainable Development Goals (SDGs) proposed by the UN are not independent of each other. In this article, we analyze the reciprocal influence of 17 SDGs, on the basis of which we propose Network-based Priority Selection Algorithm (NPSA), in which static and dynamic networks are constructed. The static network is established to describe the fundamental interaction between 17 SDGs by using Cross-impact Matrix for Targets. Fiscal expenditure, main driving force of the development of 17 SDGs, is mainly invested in ten fields. So we set up a Field-Goal Map by counting the number of targets they include in common. Then we establish four priority selection indicators, including centrality, system impact, urgency and development-based modification. According to the various developing level of different countries, we assign personalized weight vectors for them to synthesize the influence of the four indicators using the Entropy Weight Method. Then we get priority score of each goal and its rank for each year. We quantitatively simulate system dynamics to implement scenario analysis on a ten-year horizon. Once a year, we re-rank the priority of 17 SDGs and then use conservative, steady, or aggressive strategies to allocate financial expenditure. Considering the delay effect between investment and reward, we propose Delay-based Breadth-First Search Algorithm (D-BFS) to update our network, broadcasting the influence of one goal to another once a month. Logistic curve in is used to simulate the S-shaped curve change of the degree of mutual influence between goals. We use the dynamic model to give guidance on financial investment for each goal in the next ten years. Moreover, facing global field mutation and crisis, the development strategy calculated by our model has excellent stability. The code is released at https://github.com/GoodMorningPeter/NPSA.

With the continuous development of steganalysis technology and the change of network environment, the performance of traditional steganography is unsatisfactory. Due to the popularization of social networks, the use of people's behavior on social networks to achieve steganographic transmission has emerged as a hot direction in the field of information hiding. To address the drawback of low transmission security in traditional behavior steganography, this paper proposes a behavior steganography scheme in social networks based on AES encryption algorithm. The scheme uses the AES encryption algorithm to encrypt the information before transmission, which effectively enhances the security of the transmission. The experiments are performed and the results show that the proposed scheme has good feasibility and high security, and its performance is excellent.

Although deep learning based X-ray bone age detection algorithms have achieved good results, the complexity of these SOTA methods is also increasing. In this article, we propose a new lightweight network that exceeds the accuracy of the SOTA algorithm, but significantly reduces computational complexity. Our method has an average error of 0.18 years and a parameter quantity of only 0.9M, which outperforms the existing best methods in terms of computational performance and accuracy.

The Job-Shop Scheduling Problem (JSSP), one of the traditional scheduling issues in combinatorial optimization problems, has been the subject of much research for many years. In this study, to further take into account the influence of time sequence information on the operation between different processes, we develop a network architecture that combines graph neural networks and long short-term memory networks. We then train a policy network using proximal policy optimization to efficiently extract node features from JSSP disjunctive graphs and automate the learning of Priority Dispatch Rules (PDRs). The experimental results demonstrate that the proposed model for solving the JSSP problem has a high generalization performance and a global performance that is noticeably better than the current best PDRs, offering a useful approach for solving the JSSP problem.

For motion target detection in dynamic backgrounds, most traditional methods have drawbacks, such as long computation time or high limitation on the background. This paper proposes a deformable convolution-based motion target detection algorithm by replacing part of the C3 module in the YOLOv5 feature extraction layer with deformable convolution (DCNv3) to introduce long distance dependence and adaptive spatial aggregation, and adding an ECA attention mechanism to reduce the effect of background variation. The the accuracy, recall and mAP of the improved YOLOv5s algorithm increases by 1.6%, 3.8% and 2.3% respectively, and is able to identify moving targets in dynamic backgrounds more accurately than the original algorithm.

Due to the unique structure of the shelter, the cost of trial production and experimentation during the design phase is high. In order to study the statics characteristics of an information shelter, this paper discusses the structure of an information shelter, and designs and uses the ANSYS finite element analysis software to establish the equivalent finite element model of the overall structure of the information shelter, and conducts statics simulation calculation of the cabin structure under specific working conditions to obtain its stress and strain state. At the same time, the modal superposition method is used to analyze the overall transportation vibration response spectrum of the information shelter. The results indicate that the maximum stress value of the intelligence shelter is less than the allowable stress, and the maximum deformation fully meets the requirements. At the same time, it meets the adaptability requirements of the transportation vibration environment of the intelligence shelter, providing a design reference for similar intelligence shelters.

With ultra-dense networks, the large number of densely deployed low-power base stations creates more serious interference problems for the network. To address this problem, we introduce Dyna-Q learning into the power control problem of ultra-dense networks and propose a power control algorithm based on Dyna-Q learning. Firstly, we build an experience pool to store the state and selected actions of the agent during the operation; Secondly, the agent observes the current state, uses an action selection strategy to choose the next action, and then interacts with the environment to update the Q table; Finally, the agent learn again based on the virtual experience provided by the experience pool. The simulation results show that the normalized capacity of the method in this paper is improved by 0.88b/s/Hz on average and the resource consumption is reduced by 9.9% on average compared with the existing methods.

In RFID systems, the dynamic frame slot ALOHA algorithm is an efficient and low-cost method to solve the tag anticollision problem. To solve the problem of excessive energy loss of the reader caused by the Q algorithm frequently transmitting frame length adjustment commands, a dynamic frame slot ALOHA algorithm based on BP neural network (BP-DFSA) is proposed, which can estimate the remaining number of unidentified tags based on the quantity of the three slot states of the previous frame, and adjust the frame length of the next frame only at the end of the current frame. The MATLAB simulation results show that the algorithm can make the RFID system send the frame length adjustment command within 13 times in the whole identification process, and the system throughput rate is stable at about 0.345. It is calculated that this algorithm can significantly reduce the number of frame length adjustment commands sent, saving 43% of reader energy consumption compared to the Q anti-collision algorithm.

Simultaneous localization and mapping (SLAM) and path planning are key technologies for robot navigation. The grid map generated by SLAM technology is a prerequisite for the path planning algorithm of mobile robots, playing an important role in autonomous robot navigation. 2D laser SLAM relies on distance information perceived by 2D laser radar to perform pose estimation and map construction, with advantages of high mapping accuracy and stable performance. In this article, we built an experimental platform using the ROS robot operating system, simulated and validated the Gmapping algorithm, and applied the validated algorithm to the Autolabor Pro1 robot platform for indoor mapping. The results show that the algorithm requires less computation and has higher accuracy in constructing small-scale maps, with good mapping effects.

The objective of transfer learning is to learn well-established knowledge from label-rich training data to handle learning tasks on the target domain, where the labeled samples can be hardly collected and the data distribution is different from that of the training data. Nowadays, the label information in the target domain may exist in multiple different source domains, but none of which can cover all the information. To make use of the imperfect source domains, many novel transfer learning algorithms explore potential properties among them to improve the completeness of transferred knowledge. However, these works only build on the original source domains, limiting the possibility to uncover additional supplementary information. In this paper, we propose novel multi-source combined transfer learning(MSCTL). MSCTL first replaces original source domains with a new training set space which can provide richer latent information by correlating multiple source domains. Then, through a triplex-complementary strategy, MSCTL extracts and integrates threefold concepts from the space to make the knowledge complementary and integral. Furthermore, we design MSCTL as an optimization problem and provide an iterative algorithm to solve it. Finally, massive experiments are performed to verify the effectiveness of our method.

In response to the challenges of small targets and complex backgrounds in remote sensing image detection, this paper proposes an improved algorithm based on Yolov5s, named Yolov5s-RSD. The algorithm replaces the original downsampling method with SPD-Conv to reduce information loss during downsampling, adds a detection head for small targets to fully utilize their features, and introduces Biformer attention to address complex background issues. Testing on the VisDrone datasets shows that the improved algorithm achieves a 10.0% increase in map@0.5 and a 7.0% increase in map@0.5:0.95 compared to the original algorithm.

The current conventional source-load intelligent tracking algorithm of distribution network mainly realizes active power control by calculating the regulation amount of output power, which leads to poor tracking effect due to the lack of intra-day scheduling optimization of distribution network. In this regard, the fuzzy prediction-based distribution grid source-load intelligent tracking algorithm is proposed. The multi-scenario technology is used to model the power output of distribution network power devices and the power load, and build the tracking scenario model; the intra-day optimization model is built, and the MPC control method is combined to realize the control of the power output and load situation of the distribution network; finally, the power fluctuation index is introduced to characterize the source-load tracking situation. In the experiments, the power control performance of the proposed method is verified. The experimental results show that the maximum power fluctuation value is low when the proposed method is used for source-load tracking, and it has a better power control performance.

Ship route optimization is a complex multivariable multi-objective problem. To solve this problem, a feasible ship navigation optimization method is proposed, using the waypoints of the route as decision variables, the non navigable area and speed as constraints, and the travel time and fuel consumption as evaluation functions for the quality of the route. A multi-objective optimization model for travel time and fuel consumption under the influence of wind, waves, and currents is established. Then, an improved genetic algorithm with cross mutation operator adaptive evolution is designed to improve the convergence speed of the algorithm. Finally, simulation experiments were conducted on this method and compared with the data of the large circle route. The results indicate that the route planned by this method can successfully reduce flight hours and energy consumption.

The fiber optic sensing perimeter security system has broad application prospects due to its unique characteristics such as long transmission distance, wide monitoring range, high sensitivity, and strong real-time performance. Based on DWT and STFT, this paper extracts the features of the intrusion source signal and uses the CNN-LSTM deep learning algorithm to intelligently classify the signal, improving the separability of disturbance events and accurately judging the intrusion source type. The average accuracy rate of classification can reach more than 96%, with excellent intelligence and low false alarm performance. This technology is a perfect replacement for traditional perimeter alarm systems such as infrared beams and electronic fences, especially applicable for real-time measurement of vibration-related events such as long-distance, omni-directional, and multi-point illegal intrusion, illegal destruction, and structural damage. In addition, the application of distributed optical fiber sensing system in the industrial chain can improve product quality, ensure production safety, and realize production line automation.This research will improve the reliability and accuracy of the perimeter security system, improve the public security environment, save social resources, and improve social security.

To solve the problem of obstacle avoidance cars find the shortest path without collision between two locations. Firstly, the method of building a raster map is used in MATLAB to simulate the working environment of the trolley. Secondly, a hybrid algorithm combining particle swarm algorithm and genetic algorithm is proposed, and it is applied to the path optimization of the trolley. The hybrid algorithm combines the traditional particle swarm algorithm and the genetic algorithm, retains the speed and position iteration mode of the particle swarm algorithm, and integrates the cross and mutation operations in the genetic algorithm. The results show that compared with the traditional particle swarm algorithm and genetic algorithm, the hybrid algorithm can enhance the ability to jump out of local extremes in the process of population search, improve the convergence progress of the algorithm, and ensure that the obstacle avoidance trolley can find a more suitable moving path between the starting point and the target point.

The fuzzy improved DWA autonomous charging algorithm is proposed for the problems of high hardware cost, complex docking process and low docking efficiency of robot autonomous charging technology. A horizontal attitude adjustment function and a vertical attitude adjustment function have been added to the azimuth evaluation function of the traditional DWA algorithm. During the robot electrode docking process, the robot uses a fuzzy control algorithm to adjust the weight coefficients of the horizontal posture adjustment function and the weight coefficients of the vertical posture adjustment function according to the horizontal distance between itself and the charging pile, in order to control the robot's direct navigation to the charging pile.

This study examines the assessment of construction safety risks in green building risk management. By comprehensively considering all factors that affect the safety of green building construction, an evaluation index system is established. The AHP method is used to weigh each index, expert scoring determines the risk level of each project index, and the weighted summation results in the expected output value. In addition, the genetic algorithm (GA) is introduced to optimize the BP neural network, and an evaluation model based on AHP and GA-BP neural network is established that can be used to predict and evaluate construction safety risks. The results show that this method can effectively predict the safety risk of green building construction and provide valuable guidance for the safety management of green building construction.

The problems of power quality disturbance (PQD) in modern power systems are complex and diverse, while the traditional model can only extract features of a single mode, which obtains insufficient information and affects the accurate identification of PQD. In order to obtain more comprehensive information and improve classification accuracy, this paper proposes a multimodal parallel feature extraction network for power quality disturbance classification. The model uses Gated Recurrent Unit (GRU) to extract temporal features. Meanwhile, the lightweight EfficientNet (LEfficientNet) module is designed to extract spatial features and fuse multimodal features. Then output to support vector machines (SVM) for classification. The classification results of 20 PQD simulated signals show that the proposed model achieves 99.89% correct classification rate, which is not only better than the performance of a single modal model, but also has lower training cost.

Target detection tasks on water have been affected by water surface fluctuations, light changes, occlusions, overlaps and targets that are too small, resulting in poor recognition accuracy and detection speed. To solve this problem, this paper proposes an improved water target detection algorithm based on YOLOv5. Firstly, recursive gated convolution module were used instead of ordinary convolution to achieve high-order spatial interactions and improve model performance and confidence. Secondly, in order to solve the problems of difficulty in detecting small targets, the input size of the image is increased and P6 detection heads are introduced to improve the detection accuracy of small targets. Finally, to alleviate the situation of poor overlapping target detection, coordinate attention(CA) mechanism is introduced and incorporated into the backbone to extract large range features, and reduce the number of parameters and computational overhead. Compared with the original model of YOLOv5s, the improved algorithm model improves the recognition accuracy by 4.6% and the recall by 3.4%, reaching 69.2% and 67.6% respectively. The running speed reaches 208FPS, meeting the requirements of real-time detection.

Aiming at the shortcomings of traditional optimization algorithms in solving speed and flexibility of multi-UAV cooperative task planning, a multi-UAV cooperative task planning method based on improved adaptive simulated annealing and genetic algorithm is proposed. Starting from the problem of motion encounter and the starting state of UAV and target, the multi-UAV cooperative task planning model based on action sequence is constructed, and the mathematical analysis method based on task time is given. By introducing the idea of edge filling, the algorithm structure, coding, and population update strategy are improved, and an adaptive optimization algorithm is designed. The simulation results show that the above algorithm can improve the model optimization speed and task adaptability, which is suitable for the optimization of multi-UAV cooperative task planning problem.

This paper proposes a road condition detection approach based on modified YOLOv5s to address the issues of low detection accuracy and high leakage rate in the current road condition detection algorithms. Firstly, the ASPP module is introduced to not only expand the sensory field to obtain rich background information but also retain the key information of small targets in the original feature map, thus improving the detection precision in complex traffic environments and solving the problem of small target miss detection; secondly, through the localization loss function SIoU, a new angle loss is added to the penalty term to further improve the detection accuracy and convergence speed of the model. The experimental results show that the improved model achieves 84.8% detection precision, 89.1% mAP, and a detection speed of 29 frames/s. The improved model can be applied to various complex traffic environments and meet the requirements of road condition detection in the realm of intelligent driving.

Swarm intelligence algorithms have a large number of applications in solving robot path planning problems. The traditional artificial bee colony (ABC) algorithm has strong global ability, large randomness, however there are problems such as the more nodes, the higher the path complexity. In this study, node movement artificial bee colony algorithm (NMABC) is proposed overcome the disadvantages of the traditional ABC in solving robot path planning problems. The NMABC algorithm contains two novel mechanisms which are the improved initial path generation way and the neighborhood search method of the iterative path of the employed bee. The experimental results show that the NMABC has good global search capability and fast iterative convergence, and it can find the optimal path quickly.

With the development of deep learning, small target detection application scenarios are gradually becoming more widespread, and higher demands are made on small target detection accuracy. The attention mechanism is a simple and convenient way to improve target detection accuracy. In order to improve the detection performance of the attention mechanism for small targets, a SimAM-based attention method for small target detection is proposed by us, which changes the activation method of SimAM and add hyperparameters to improve the detection effect of small targets. The experimental results show that the algorithm has good detection performance.

This paper proposes a method to better address the optimization problem of cable forces in cable-stayed bridges, particularly when considering multiple objectives simultaneously. It applies genetic algorithms to optimize cable forces and utilizes Pareto theory for multi-objective optimization. The proposed method is implemented using MATLAB software as the programming platform and ANSYS parameterized modeling. Additionally, an actual bridge is used as a case study to validate the proposed optimization method. By minimizing the main beam bending strain energy and the maximum cable force of the cable as objectives with cable force values as optimization variables, the internal force state of the structure is improved by adjusting the cable forces distribution. The Pareto optimal solution is obtained through optimization calculations, reflecting the minimum main beam bending energy that can be achieved under different cable force safety factors. This solution can guide engineering design.

Using a regional integrated energy system with the advantages of multi-energy complementarity and coupling, it can comprehensively enhance the optimal allocation of multiple energy consumption. This paper proposes a multi-objective low-carbon optimal dispatch model for a regional integrated energy system , which fully exploits the advantages of heat storage devices, electric boilers and electric energy storage in regional energy systems to promote the grid-connected consumption of wind power. By establishing a load elastic demand response mechanism, the peak-to-valley difference level of system load is reduced to further enhance the grid access space of wind power. Finally, the proposed scheduling model is optimally solved using the NSGA-II algorithm under the improved elite strategy.

Federated learning is a promising new paradigm for privacy protection. By dispersing training data storage locally rather than centrally constructing large-scale datasets, federated learning provides a mechanism for training models that protects privacy for various data-hungry applications, such as recommendation algorithms for mitigating information overload. This article implements a family of algorithms with common characteristics in a federated learning framework for model parameter transfer. Through comparative experiments, the performance differences of these algorithms under this federated learning framework are examined, and their model performance is compared and analyzed under their respective centralized learning forms. The conclusion is that the algorithm family combined with neural networks has higher performance and is more stable in this federated learning framework, with less performance loss.

Since 2020, the national power situation has been very tense. With the national "carbon peak" and "carbon neutral" goals proposed, and the gradual establishment of a green, low-carbon and environmentally friendly economic system, if the future power load of the supercomputing center can be accurately predicted the situation, and according to the power dispatch and regulation, can achieve the purpose of energy saving and emission reduction. Power load forecasting using statistical and artificial intelligence technologies still has a lot of issues today. In order to increase the accuracy and stability of power load adjustment, this research suggests a novel Transformer approach based on the attention mechanism. It establishes a power load analysis and prediction model. This paper can provide the supercomputing center with the load demand situation in the future, and it has important application value for the regional power management department to formulate development plans and effectively manage the load, and is of great significance for responding to the national energy conservation and emission reduction policy.

Wildlife protection is a critical issue in natural conservation, and it is necessary to establish a scientific model to provide more living places for wildlife while balancing human interests. In this paper, we propose a Markov state transition model to optimize the management mode of different areas in a wildlife reserve and use genetic algorithms to simulate future trends and compare management models. Our results suggest that the reserve should be divided into two areas, one for the survival of wild animals and the other for human visits, with the survival area covering more than 70% of the reserve's area. Our prediction model shows a 30% growth trend for wild animals in the survival area and a 20% change trend in the human visit area, indicating the effectiveness of our approach. Finally, we provide specific strategies to achieve a comprehensive and balanced approach to human-wildlife conflict.

As a common measuring instrument, water meters are widely used in places such as water and power plants and households. The manual reading mode of old-fashioned water meters requires a large amount of manual and time costs. With the continuous breakthroughs in deep learning theory, it has become possible to use convolutional neural networks to automatically read old-fashioned meters. In order to improve the accuracy of meter reading and exclude other interference factors from the dial, this article proposes an attention-based DB network to make the model pay more attention to the meter display box, and uses a CRNN network for water meter reading, which effectively improves recognition accuracy. The model was tested and evaluated on a dataset of old-fashioned water and power plant meter dials, and the experimental results show good performance in the recognition accuracy of single and whole string characters. In addition, we deployed the model on Aiot to implement a high-precision and stable meter reading system.

Convolution neural network is widely used in various fields. The convolution layer is the core layer of the convolution neural network. The back propagation speed of the convolution layer will directly affect the training speed of the whole network, thus affecting the whole performance. For the convolution layer with stride ≥ 2, the error transmission phase of back propagation will carry out a large amount of padding in the feature graph, resulting in a large amount of additional overhead in access and calculation. In this case, we propose a new optimization method, which can reduce the overhead caused by padding to almost zero, and implement it by implicit convolution on domestic heterogeneous platforms. The experiment shows that the performance of the operator optimized by this method is nearly 50% higher than that of the original operator of the platform, and the average performance reaches 90% of that of NVIDIA V100 operator.

Water is an important resource for human survival. It has become an urgent problem for people to obtain large-scale and real-time water change information. The surface water source area of Hunchun city is selected as the study area. Combined with the field measured data and quasi synchronous landsat8 data, the water quality inversion models of total nitrogen, total phosphorus and chlorophyll a in the study area are established based on BP neural network and improved PSO-BP neural network. Nine test points are randomly selected to evaluate the accuracy of the two models.The results show that: (1) The precision of improved PSO-BP neural network model is better than that of BP neural network. (2) In general, the water quality concentration of Hunchun surface water source area is different. (3) The total nitrogen in Hunchun surface water source is mainly Class II water quality, the total phosphorus is mainly Class IV water quality, and the chlorophyll a is Class I and Class II water quality. It can be seen that the inversion effect of surface water quality parameters based on improved PSO-BP neural network technology is better than that of BP neural network.

Named entity recognition (NER) involves two main types: nested NER and flat NER. The span-based approach classifies entity types by head-tail pair span representations and can handle nested and flat entities uniformly. However, the span-based approach uses a single feature and ignores the relative position information between head-tail pairs, which affects the precision of entity recognition. Therefore, we propose a nested entity recognition model that combines rotary position embedding and biaffine attention mechanism (RoPE-BAM) to improve the model performance by adding relative position features to the span representations. Concretely, we first obtain the head sequence and tail sequence representations through two feedforward networks. Then, to incorporate the relative position features, rotary position embedding is applied to both head and tail sequences. Finally, we use a biaffine attention mechanism to capture the span representations while generating the relative position information in the span. Extensive experiments were conducted on five widely-used benchmark datasets to demonstrate the effectiveness of our proposed RoPE-BAM model.

In this paper, we introduce a novel method for the classification of ancient glassware, combining feature selection using the Gini coefficient, spectral clustering, and Bayesian-optimized XGBoost. Ancient glassware, a significant cultural heritage, provides insights into the history and technology of bygone civilizations. However, the classification of such artifacts presents a formidable challenge due to the lack of standardized criteria and limited expert availability. Our proposed approach addresses these issues by identifying key chemical composition features using the Gini coefficient, grouping similar samples using spectral clustering, and classifying them into their respective categories utilizing Bayesian-optimized XGBoost. We demonstrate the efficacy of our approach through experiments conducted on a dataset of ancient glassware.

Due to the lack of sufficient training samples, it is not possible to accurately identify the visual relationships of concrete continuous beam bridge construction scenes, resulting in a low detection recall rate. Therefore, a new visual relationship detection method for building scenes is proposed. Perform feature fusion and optimization processing on the construction scene, extract unique features and spatial distribution features of the construction scene, use an improved ant lion algorithm to simulate the walking motion of ants searching for food, establish a matrix, obtain local optimal solutions, and then detect the visual relationship of the construction scene. Experimental analysis shows that after the application of the new method, the recall rate of visual relationship detection has significantly improved, reaching over 96%.

In order to solve the problem of lack of temporal and multi-scale information in gait recognition tasks based on gait energy images. A gait recognition method based on two-branch network is proposed. In the convolutional network branch, the attention mechanism and residual connection are combined to design a multi-scale feature extraction module to obtain effective multi-scale features. A simple temporal image is constructed and input into the long short-term memory network to extract temporal features. More accurate gait classification can be achieved by combining the two types of features. Experimental results on open gait dataset CASIA-B show that the proposed method has good classification effect.

With the continuous improvement of artificial intelligence technology and the rapid development of big data applications, deep learning technology has been widely used in various fields. Although deep learning has achieved great success in many fields, it still faces some potential problems. For example, the insecurity of neural networks has raised concerns as this may lead to security vulnerabilities in critical areas. This article proposes an adversarial attack method based on unsupervised algorithms to search for sensitive points in time series dataset and use differential evolution algorithm to perform one-point attack on sensitive points. The experiment was conducted under various experimental parameters. All experimental results are presented in the form of a heatmap, and the article provides a detailed analysis of the experimental results.

Legal judgement prediction (LJP) means to predict the judgement results according to the given facts of the cases. Charge Prediction (CP) as a sub-task for LJP, is a substantial research area in Judicial field. The key problems remain in CP task includes: (a). low-accuracy on low-frequency charges and similar charges; (b). the lack of interpretable model which are vital in judicial appliance. To address the aforementioned problems, we propose a CP model based on keyword extraction and tree structure, e. g. Tree-CP. By fusing the keywords into word embedded layer and deploying the concept tree in Attention layer, the interpretability of the predictions is solidly enhanced. Tree-CP leverages various data augmentation method on low frequency charges to amend the small sample pool. The features are extracted and fused via BiGRU and capsule network separately to improve the predicting accuracy. The comparative experiment proves the effectiveness of Tree-CP. And the ablation experiment shows the soundness of the model.

As the scale of transmission lines is expanding, the safety of transmission line operation is getting more and more attention; insulators as an important component of transmission lines, the use of insulators greatly affects the safe operation of transmission lines. The traditional depth recognition algorithm can not achieve effective recognition of infrared images, in order to achieve rapid recognition of insulators in complex environments, this paper proposes the recognition of infrared insulator images based on the improved YOLOv5 depth neural network detection algorithm. Firstly, Ghost convolution was introduced into the backbone network to speed up detection and network lightweighting; secondly, to enhance the multi-scale convergence of networks, improved GAM attention module added behind the neck network; in addition, the network introduces an EIOU loss function for convergence; finally, validation of this improved algorithm with the collected infrared insulator dataset. The results show that the improved algorithm in this paper achieves 91.2% accuracy and 92.1% mAP on the infrared insulator dataset, which improves 2.3% and 4.1% compared with the test results of YOLOv5 model, simultaneous detection speed up to 91 FPS, which meets the real-time requirement.

With the continuous expansion of urban transportation networks and the increase in the number of vehicles, real-time monitoring and identification of urban transportation networks has become an important research direction. This article proposes a real-time super-resolution image recognition method based on large-scale urban vehicle network interaction. This method improves the efficiency of traffic management and control by monitoring and identifying vehicles in the urban transportation network in real time. This article adopts deep learning and image super-resolution techniques, and uses large-scale datasets for experimental verification, achieving good results.

The research of TCM informatization is usually based on the text of the patient's symptoms. TCM physicians use synonyms to describe a patient's symptoms, which can cause data redundancy in the patient's symptom text. Therefore, in the process of TCM informatization, it is first necessary to process the TCM text and standardize the synonyms in the symptom text. In this paper, a synonym set is constructed based on the symptom text of patients, and the ED-BERT model is proposed for the normalization of symptom texts in traditional Chinese medicine, which lays a foundation for the follow-up research of traditional Chinese medicine informatization. The accuracy of the proposed model for TCM synonym recognition reaches 0.8613, which has advanced advantages over traditional methods.

Deep Interest Networks (DINs) capture potential user interests by mining long-term and short-term behavioral sequences of users, thereby predicting the information that users are interested in. However, vanilla DINs are unable to handle user behavior sequences exceeding a predetermined length, resulting in the inability to effectively capture potential long-term needs of users. To address this issue, an Attention-based Interest Network(AIN) is proposed which utilizes aggregated user behavior sequences. AIN aggregates users' historical behavior sequences and adopts attention mechanism to enhance expression capabilities. Experimental results demonstrate that AIN can significantly improve recommendation performance. Compared to vanilla DIN models, AIN achieves the best AUC of 0.6429 and ReIaImpr of 7.44.

The segmentation of potato disease leaves is of great significance to improve the accuracy of disease identification and reduce the loss of potato yield. Considering the single instance category in the image segmentation task of potato disease leaves, the classification branch in Mask-RCNN model is removed and the color matching branch is added. The K-means clustering algorithm was used to extract the RGB values and proportions of the five maximum colors of the leaf image segmented by the model. Then input the RGB values and proportions into the color matching loss function to calculate the color matching loss. Finally, the color matching loss is fused with the Mask-RCNN model’s bounding box regression loss and mask loss to form a joint loss to guide the parameter update and optimization of the entire image segmentation model. The average intersection ratio of the improved Mask-RCNN model image segmentation reaches 93.53%, which is significantly improved compared with that before the improvement. It shows that proposed method can improve the precision of the segmentation of diseased leaves, and has practical significance for the increase of potato yield.

Due to the lack of pixel level structural matching data, thermal infrared grayscale images are more difficult to color than visible and near-infrared grayscale images. Therefore, this paper proposes a unsupervised learning method based on CycleGAN. On the basis of CycleGAN, a pre trained edge monitor is introduced to calculate the edge feature map before and after image transformation, and the edge similarity loss function is calculated as the basis for optimizing the neural network parameters. The experimental results show that the proposed method effectively reduces the loss of effective edge information during the coloring process and suppresses the generation of abnormal edge information during the coloring process.

The development of the Internet has led to a sharp increase in the amount of Chinese text data, and the classification of short Chinese texts has problems such as fewer words, more ambiguity, sparse features, and irregular features, which leads to poor results of traditional text classification techniques. This paper designs and implements the ERNIE-TEXTCNN model. The model uses ERNIE as the word vector model, TEXTCNN extracts features from the input text vector, and softmax obtains the final classification result. Through comparative experiments with other methods, it is shown that the ERNIE-TEXTCNN model can extract text feature information very well, and has a good classification effect of Chinese short texts.

In order to quickly and accurately cognize targets and formulate optimal strategic solutions, this paper conducts a study of a comprehensive analysis model system for target values. The system is based on target intelligence information and actual requirements, and aims to efficiently assess the value of each entity target. Based on complex network theory, this method topologically abstracts the target types in the territory into six types of nodes and five types of links considering the influence of objective factors, and constructs a system model combining the subjective factors of commanders' decision making. And then the subjective willingness of commanders is fused through expert evaluation, using gray correlation to generate subjective decision indicators, combined with objective network attributes to form a combined subjective and objective target value analysis method. Finally, the complete EGE algorithm is constructed by combining the entropy value method, and the effectiveness and feasibility of the algorithm were verified by model simulation and comparison with traditional methods. The results show that the method is more in line with the actual mission evaluation requirements and is more adaptable to the actual mission scenarios.

In this paper, a Deberta-BiLSTM-CRF model based on the fusion of character and lexical features is proposed to solve the problems of semantic lack and entity nesting in Chinese medical texts in entity recognition tasks. The model uses the Deberta pre-training model and the Soft-lexicon method to dynamically fuse dictionary information and enhance the semantic representation of text. After the learning and decoding of BiLSTM and CRF models, the effective recognition of medical entities is realized. Follow-up work will consider making a large-scale Chinese medical dictionary to improve the recognition effect of the model. The experimental results show that the model proposed in this paper has a greater improvement in the F1 value of entity recognition than the RoBERTa-wwm-BiLSTM-CRF and Bert-BiLSTM-CRF models, and the speed has also improved. Therefore, the model has broad application prospects in Chinese medical entity recognition.

Existing robotic arm control systems usually rely on artificial inverse kinematics modeling to solve, but due to the complexity of the robotic arm, this method is often expensive, particularly when the structure of the robotic arm changes, requiring it to be remodeled and solved again, making it too complex. This paper proposes a neural reinforcement learning-based inverse kinematics robotic arm control system, which only needs to model the forward kinematics of the robotic arm in the virtual simulation environment to JIT generate self-supervised training data to train the neural network, and the network can be used to control any size of hardware robotic arm after the network training converges, as long as its joint arm length ratio is the same as the ratio of training. Experiments show that we can complete this task with a small neural network, and low latency inference servo angle parameters under $150 Jetson Nano, in virtual simulation testing, our control accuracy reaches 97%, while in hardware testing, the accuracy of the camera and servo is reduced, but the control is still effective in the general range. Therefore, this inspires us to study the future research on servo angle nonlinear error correction and the end position control technology under the controlled environment.

This paper aims to apply sentiment analysis techniques to cope with the rapid growth of public opinion data. Due to the complexity of Chinese semantics and the sparsity and high-dimensionality of text data, sentiment analysis faces great challenges. To this end, this paper proposes a deep learning model based on RoBERTa. The model uses the RoBERTa model to encode text into word vectors, and weights the word vectors with a weight matrix to enhance the emotional features in the sentence. Then, the contextual features of the word vectors are extracted by BiLSTM, and the hidden feature vectors are weighted by the attention mechanism, and finally classified by the fully connected layer. Experiments show that compared with the BERT model and other text classification models, the accuracy of the model proposed in this paper is significantly improved in the three public data sets.

Gearbox is a key component of large rotating machinery. To ensure the safe operation of large rotating machinery and to meet the demand for automatic and intelligent gearbox fault diagnosis, this paper investigates the use of the frequency spectrum of gearbox vibration signals in an improved One-dimensional Convolution Neural Network (1D-CNN) to diagnose early weak faults in gearboxes. In view of the problem that the maximum pooling algorithm loses too much information and the average pooling algorithm blurs important information in the common convolution neural network (CNN), a small-scale convolution kernel with a moving step of 3 is proposed to replace pooling operation, so as to realize the role of automatically adjusting the weight extraction characteristics in the process of down-sampling and model training. Also, the global max pooling is used to extract the effective information of the final convolution layer instead of the global average pooling, and the regularized Dropout technique is used to further optimize the deep convolution neural network (DCNN), significantly reducing the number of network parameters and minimizing the phenomenon of network over-fitting. The research is carried out on an experimental platform for gearbox failures, the results show that the improved 1D-CNN has a correct fault identification rate of over 98%, which is 20% higher than that of the previous CNN, and the network has better noise immunity and stability. The method in this paper can provide a basis for researchers in the construction of DCNN and has certain practical engineering application value for fault diagnosis of gearboxes.

In order to solve the reactive power optimization problem of distribution network with distributed generators, this paper proposes an improved driving training-based optimization algorithm. Firstly, the reactive power optimization model is established with the reactive power output of distributed generators and static reactive power compensator as the control variables, and the minimum active power network loss, reactive power compensation cost and voltage overlimit deviation as the optimization objectives. Then the basic driving training-based optimization algorithm is improved by introducing good point set, golden sine algorithm, Gaussian variation and chaos disturbance. Finally, the enhanced IEEE33 bus system is taken as an example to verify and analyze the proposed reactive power optimization model and improved algorithm. The experimental results show that the proposed algorithm is effective in solving the reactive power optimization problem of distribution network with distributed generators.

In order to solve the problems of complex and time-consuming traditional manual feature engineering methods in the field of human activity recognition, incomplete grasp of local information by a single recurrent neural network (RNN) and lack of utilization of temporal information by a single convolutional neural network (CNN), this paper uses the Gramian Angular Field (GAF) algorithm to convert inertial sensor data into images and designs a multimodal fusion based on CBAM-CNN-Attention method, which can effectively improve the human activity recognition accuracy. Firstly, the time series and image features are extracted using one-dimensional convolutional kernel and two-dimensional convolutional kernel, respectively. Secondly, the most important features are obtained and optimized using the attention mechanism. Finally, the recognition results are classified using a softmax classifier. The experimental results on the UCI-HAR dataset show that our proposed method has superior performance to other methods that use only data collected by accelerometers.

Sentiment analysis has always been a hot research topic in the field of natural language processing. With the development of artificial intelligence and big data technology, many data such as text, images, and videos that have personal emotional tendencies are accidentally generated. Initially, researchers mainly focused on studying text data, but over time, more and more people began to realize the limitations of single-mode analysis. By introducing multi-modal data such as images and videos, more dimensional information can be provided for sentiment analysis, thereby improving performance. Therefore, more and more researchers are exploring richer sources of emotional information to improve the effectiveness of sentiment analysis. This article proposes a multi-modal sentiment analysis method based on deep learning and designs an experimental model targeted towards video modal data. In terms of text, this article proposes a GloVe-based BiGRU word vector model to process data. For images, the method changes from using simple frame extraction to speaker anchoring, uses object detection technology to segment the region of interest for the subsequent model to extract image expression features. A pre-trained ResNet101 model is used to obtain vectors and generate a sequenced image matrix, which is then input to LSTM for processing. From the experimental results, it can be seen that this method performs better than existing methods in terms of accuracy and F1 scores.

Visible-Infrared person re-identification technology aims to match the target persons across the visible and infrared modalities. In this paper, we propose a visible-infrared person re-identification method based on a modal-identity dual-central loss. Modal-identity dual-central loss constrains the network to extract modal shared features by pulling in the infrared modal center and visible modal center of the same identity person, while pushing away the identity centers of different person to maintain inter-class discriminability. In addition, to extract more discriminative information, we propose a feature pyramid integration network based on efficient channel attention. Specifically, the network fuses high-level features and fine-grained low-level features to build a multi-scale feature map, and introduces an efficient channel attention module to enhance the salient features of person. Extensive experiments have been conducted to validate our proposed method on the SYSU-MM01 and RegDB datasets.

Deep unsupervised domain adaptation (UDA) methods are widely used in the computer vision. However, such methods are usually negatively affected by inaccurate pseudo-label, which impairs model performance. This paper proposes a new UDA framework for optimizing pseudo-label by using the mean teacher model (LOOT). LOOT first constructs a teacher model using the source domain data to provide robust pseudo-label for subsequent training. A threshold strategy is also proposed to drive the model to learn to a more accurate output. Second, an optimal transport method using the JSD regular is proposed to align the global distribution of the source and target domains. In addition, the model is equipped with clustering objectives to enable discriminative domain invariant feature learning. The combined experiments show that LOOT achieves state-of-the-art result in the benchmark datasets.

In clinical practice, deep learning models commonly encounter performance degradation in actual applications, mainly due to domain shifts caused by heterogeneity from various sources such as scanners, stains, and medical sites. To address this issue, we examine the single-source domain generalization problem - developing a deep network that can withstand unseen domains, using training data from only one source domain. Our solution involves a novel domain-invariant feature learning model, integrating attention normalization and style enhancement modules at shallow layers of the network while exploring the optimal combination strategy in the convolution block. Furthermore, we use color transformation technology to randomly convert images to produce images with different color distributions to eliminate the effects of stains. The experiments demonstrate that our proposed approach achieves state-of-the-art performance in pathological classification tasks and can be extended to unseen domains.

Entity linking task aims to establish unambiguous links between entity mentions in text and target entities in the knowledge base. Traditional entity linking models rely on complete context information in long texts. However, with the fast-growing amount of internet data, most of the data is in short text which presents a range of challenges including poor contextual semantics and disordered word order. Additionally, compared to English, Chinese poses greater challenges for entity linking task due to its unique language characteristics. To address this, we propose a multi-task learning based entity linking model for Chinese short text. By utilizing the fine-grained entity typing task as an auxiliary task for entity linking, and sharing the BERT-BiLSTM encoder, the model can leverage the implicit relationship between the two tasks to maximize potential information in the text and enhance performance in entity linking. Experimental results on Chinese short text datasets demonstrate that our proposed model outperforms a range of deep-learning based entity linking models, as well as the single-task entity linking model with the same structure, fully verifying the effectiveness of the proposed approach.

Fabric defect segmentation is an important part of ensuring the quality of product production. Using fabrics with surface defects will affect the quality and reputation of their products. In previous studies, some model compression methods have helped semantic segmentation models to be deployed on resource-limited working devices. However, the capacity reduction of models usually leads to a decline in detection performance. We propose a knowledge distillation method combining traditional KD loss and contrastive relational distillation, which makes student models learn the differential representation among various defects while receiving knowledge transfer from the teacher model. We use DeepLabV3+ and PSPNet with MobileNetV2 backbone networks as student models to validate our method. Experiment shows that our method performs better than direct training and traditional knowledge distillation methods on the DAGM and AITEX datasets. Our method enables lightweight models to achieve higher performance on fabric defect segmentation tasks.

Monitoring passenger flow distribution in real-time in subway stations plays a key role in large passenger flow control and intelligent passenger flow management. A binocular vision based method for generating heat map of passenger flow distribution in subway stations is proposed in this work, based on the extraction and analysis of the monitoring video images in subway stations. Firstly, the binocular camera is used to collect the depth information of the passenger flow in the subway station. Secondly, a Yolov5-based pedestrian detection algorithm is applied to obtain the passenger flow information and coordinate positions from the image data. Then after coordinate conversion, the actual position information of passengers is obtained. Finally, based on passenger location coordinates, the heat map of passenger flow density can be generated to monitor passenger flow in real-time. The experimental results show that the average accuracy of passenger detection is 95%, which is 2.1% higher than before the improvement. ing the passenger’s coordinates in the camera coordinate system to the actual coordinates, the resulting error at 5m, 10m, and 15m is 0.4m, 0.9m, and 1.3m, respectively. and the presentation of the heatmap is consistent with the actual situation. The research results are applicable to real-time passenger flow monitoring and statistics in subway systems, enabling operators to have a real-time understanding of passenger flow distribution and timely guidance, and have certain application prospects.

In human communication, besides direct verbal speech, facial expressions are also the main method to convey inner thoughts and emotions. By analyzing facial expressions, it is possible to obtain information about human inner emotions. Applying expression analysis algorithms in social robots can help robots accurately understand users' emotions and intentions, which in turn leads to better human-computer interaction.Therefore, in this paper, an effective & efficient expression recognition method is designed. The network uses the Ghost Module as the core module, and the lightweight attention module is used to increase the accuracy of expression recognition. In addition, the network is trained with a distributed label loss in order to solve the problems of insufficient data and long-tail of the facial expression dataset. Experiments prove that the network is superior, performs well on the RAF-DB dataset, with 86.8% accuracy without pretraining and has a fast processing speed. We also designed a real-time expression analysis system with this network as the backbone to simulate the actual working scenario of the robot, and found superior results to satisfy the functions of real-time human-computer interaction and efficient expression recognition.

In the existing person re-identification research, most of the studies are based on the person re-identification of visible light images, and the illumination conditions can affect the experimental results. Therefore, cross-modal person re-identification research has emerged, and the most commonly used method is visible-infrared person re-identification. Due to different cameras, different camera angles, different scenes, and other factors, cross-modal person re-identification faces great challenges of cross-modal differences and intra-class variations. Therefore, Mask-guided Two-stream Attention Network (MTAN) with modality alignment for infrared-visible person re-identification is proposed in this paper. First, the mask-guided two-stream feature extraction network is proposed to extract shape-aware modality features. Then, the modality alignment module is designed to eliminate cross-modal feature distribution bias. Overall, the proposed method combines the mask-guided two-stream feature extraction network, modality alignment module and the loss functions, to effectively extract modality-aware distinguishable features and reduce cross-modal distribution bias.

Short-term prediction of ocean significant wave height is crucial for human maritime activities. However, historical wave height data possess characteristics such as high complexity and strong randomness. To address these issues, this paper proposes a hybrid model that combines Ensemble Empirical Mode Decomposition (EEMD) and Generative Adversarial Networks (GANs). The model first decomposes the raw wave height data using EEMD to select effective sub-series for reconstruction and implements denoising of the original data. Then, the reconstructed wave height sequence and relevant features are fed into a GAN-based prediction model. The generator in the model comprises of Long Short-Term Memory Network (LSTM) networks, while the discriminator consists of Convolutional Neural Networks (CNNs). Finally, the generator and discriminator are trained adversarially using North Pacific oceanographic data to achieve short-term predictions of ocean wave height. This proposed model is then compared with three baseline models, and the results show that our model performs the best in all three evaluation metrics, making it a valuable tool for ocean engineering.

In response to the less referees in non -professional basketball events, this article will design a multi -functional basketball technology station based on deep learning gesture recognition and image -based car -based cars. While tracking the referee position, the corresponding gesture can reduce the labor costs in basketball games and improve the smoothness of the game. This article uses the YOLOv5 algorithm, which can reaches more than 90%of gesture recognition. Using the OpenCV image processing module can better capture the position of the referee and make gesture recognition more accurate and fast.

It is a challenging problem to identify new categories with few support samples. In the current few-shot learning tasks, it is usually assumed that the support set is accurate, but in real scenarios, the support set may still have noisy samples. Therefore, how to learn a reliable model on the noisy support set is the direction of current research. To address the situation in few-shot learning task where the support set contains noisy samples, we propose a Transformer model for noisy few-shot learning that considers both global and local features. Specifically, we consider the difference between local features and global features between noise samples and support set samples, then align the local features of the samples, and use the attention mechanism of the transformer model to reduce the interference of noise samples to the model. To evaluate the effectiveness of our proposed models, we extensively test these methods on noisy versions of MiniImageNet and TieredImageNet. Our results show that the model performs competitively in the presence of label noise.

As a fundamental task in the field of NLP, Chinese NER has attracted many researchers. Recently, the lattice structure supported by the auxiliary lexicon has been replaced by a new flat-lattice structure to accommodate the input of the Transformer encoder. The positional information in the flat-lattice is regarded as an information flow input into the calculation of self-attention. Due to the addition of this information flow, a variant of self-attention is used here. Although the acquisition of the distance feature relies on the positions of two tokens represented as characters or words in the flat-lattice, the current method only considers the interaction of the distance feature on the former token, rather than taking both tokens into account. In this paper, we propose a new variant of self-attention, which not only considers the influence of the distance feature on the former token but also considers the influence of the distance feature on the latter token. Experiments on two open datasets show that our proposed model outperforms other state-of-the-art models.

The encoder-decoder framework underpins the majority of Chinese abstract generative models. Although the generated summaries are comparable to the source text, they invariably suffer from generating inaccurate or redundant words, which can result in imprecise content or insufficient summaries. To address this issue, we propose a Roberta-based model with noise co-attention. In our way, we propose a novel noise co-attention to correct and remove the learned biased and redundant information from the source text. We first use the pre-trained model ROBERTA to obtain high-quality word vector representations and then build UNILM based on the first step, in particular adding noise co-attention layer to guide the summary generation process. Related experiments are being carried out on the large Chinese abstract dataset LCSTS, using ROUGE as the evaluation metric. According to the results, our invention surpasses numerous cutting-edge models for the summary generation and offers greater improvements over baseline methods.

Recommendation systems are the latest solution to solve the information surplus in the Internet era. However, the mainstream recommendation models have the problem of data sparsity and lack of diversity. Therefore, this paper proposes a diversity recommendation algorithm based on multi-graph collaboration and graph convolution neural networks (GCN) to solve the above problems. Specifically, we use the user-item bipartite graph to construct a multi-graph to mine the edge information of potential space, so as to improve the data sparsity problem and construct a diverse node neighbourhood, and we use the graph convolution neural networks to model the implicit characteristics of users and items, so as to learn users' diversity preferences. We conducted a series of experiments on real data sets, and the experimental results show the validity of our method (DR-MGCN).

Image classification methods are extremely devised to dispose the predictions of images labels through extracting the features. Existing methods mainly utilize unsupervised deep learning method, which will cause numerous computation costs and require human to label the images data set. Therefore, we propose a novel self-supervised deep learning method to solve the image classification issue. Initially, we extract the feature vectors of input data through utilizing a residual network. Subsequently, the self-supervised framework is established with a loss function to optimize the training results. At last, we simulate our proposed method on CIFAR-10 and MNIST classification tasks data set to evaluate the performance and compare our method with existing unsupervised learning methods. From our continuous experimental results and comparison simulations, we can conclude that our devised method can achieve the classification tasks for various input images with reasonable computation costs.

This study was undertaken in Nanchang City with the aim of enhancing resilience against droughts, which have become a growing concern and have inflicted significant harm to both the economy and the quality of life of the city's inhabitants. The research efforts utilized the monthly data of precipitation and average temperature in Nanchang from January 1960 to December 2018 to accomplish the following objectives: assess the trend and abrupt variations of precipitation and temperature via a combination of linear regression, M-K test, and sliding t-test; analyze the drought intensity and frequency through two drought indices, using different time scales; and employ NAR Artificial Neural Network to predict the SPI-1 and SPEI-1 indices, and evaluate their accuracy. The findings indicated that there was a significant increase in annual precipitation and average temperature in Nanchang over a period of 59 years. At the annual scale, droughts were more frequent and light to medium droughts were the most prevalent, whereas heavy and exceptional droughts were less frequent. At the seasonal scale, summer and autumn were more susceptible to droughts than spring and winter. On the whole, the SPEI index was found to be more effective in identifying droughts in Nanchang, being more sensitive to the variations of drought and capable of offering varying assessments of drought intensity at different time scales. The overall performance of the prediction model was deemed satisfactory, with the prediction accuracy of SPI-1 and SPEI-1 indices being 69.4% and 81.5% respectively in the short-term prediction period. This resulted in the conclusion that the model had a superior prediction effect on the SPEI-1 index. The findings of the research could serve as a reference for future predictions of drought trends in Nanchang in the short term.

In the original Informer model, an encoder and a decoder are used to process the entire time series. Although this structure is simple, its computational cost and memory requirements are very high for long sequences or high-dimensional data, making it difficult to perform effective training and prediction. In addition, since there may be large differences in the features of different time periods in the sequence, it may be difficult for a single encoder to adequately model them, resulting in a drop in predictive performance. Therefore, in this paper, by adopting multiple encoders to process the features of different time periods separately, the local structure of the sequence can be more fully utilized, thereby improving the accuracy and robustness of the model. In addition, multiple encoders can process different parts of the sequence in parallel, thus speeding up the training and prediction process and improving the efficiency of the model. In experiments, we used a set of time series datasets to evaluate our model. The results show that compared with the original Informer model, our multi-encoder single-decoder model can significantly improve the prediction accuracy and achieve higher speed under the same computing resources.

Online social network has become an indispensable place for People's Daily life and social communication. Billions of people around the world are more or less active on online social networks every day. While it brings us the convenience of real-time communication and obtaining information, it also hides the threat of fake users in the network. These fake users use malicious comments and spread fake links,seriously affecting the normal experience of users in social networks. The article proposed a fake account detection model based on node2vec and node feature merging that can help accurately identify fake users in online social networks. Experiments showed that the model proposed in this paper has some improvements in accuracy, precision, recall and F1 value compared with other models in fake account detection tasks.

In recent years, there are more and more researches in the field of video compression based on deep learning. Most of these researches maintain the same framework, including motion estimation and compensation modules. We propose a different method, which does not need an explicit motion estimation module. Instead, video compression is achieved by extracting and compressing the residuals between two adjacent frames. The residual is extracted by a 3D pyramid network. Based on the robust representation ability of deep features in various applications, the whole coding and compression process is transferred from pixel space to feature space to improve compression efficiency. The transformation process is realized by two networks: feature extraction and frame reconstruction. Our video compression framework has a simple structure and a smaller number of parameters than most subsequent proposed models. The experiment result shows that the performance of our PSNR model is better than H.264, and is comparable to H.265 in some cases, while our MS-SSIM model is better than H.265 in all cases.

Aspect Sentiment Classification (ASC) is aimed at determining the sentiment disposition towards a specific target. Most previous approaches use RNN and attention mechanisms to model context and target words. However, these approaches ignore the syntactic relationship between the aspect terms and the corresponding context words, resulting in the models wrongly focusing on syntax-independent words. To solve this problem, we propose an attention fusion model based on a Graph Convolutional Network (GCN) that considers syntactic and semantic information of sentences as well as captures more comprehensive information using a Feature Interactive Learning mechanism. We also apply the Pre-training model Bert to this task and obtain new state-of-the-art results. Experiments and analysis demonstrate the efficacy of the model.

The intelligent reading of industrial radiographic film defect information for welds has always been an important issue, usually divided into three steps: image pre-processing, feature extraction, and defect recognition. In the image pre-processing process, it is necessary to extract the weld area, reduce the number of parameters, and avoid interference from the base material area and lead identification on defect recognition. However, on most digital films after grayscale transformation, there is usually a grayscale gradient from white to grey to black from the weld to the base material, which causes the weld boundary to be blurred and the weld area cannot be directly extracted. In this article, we divided the digital films into three regions, "welding seam region", "uncertain region", and "non-welding seam region", based on the grey value changes, and assigned labels "correct", "unknown", and "wrong" to them respectively. We trained these three regions using the U²Net, and the regions predicted as "unknown" and "correct" were combined as welding seam region. The experimental results showed that the proposed method achieved a reasonable prediction rate (RP) of 95.45% on the test set, with a minimum MAE index of 0.0028.

In order to improve the function of smart health data analysis platform, this paper puts forward the application research of k-means algorithm in the design of smart health data analysis platform. A smart health service platform based on data mining is designed, and the platform is divided into three parts: data acquisition module, data storage module and data mining module. After that, the function and implementation method of each module are introduced in detail. Among them, the data acquisition module realizes the acquisition of data of Bluetooth electronic blood pressure meter; The data storage module adopts the combination of traditional database storage and cloud storage to solve the storage problem of massive health data; The data mining module uses the improved k-means algorithm and GM (1,1) grey prediction algorithm to analyze and predict the preprocessed data.

In recommendation systems, the use of information from knowledge graphs to enhance the effectiveness of recommendations has become a hot research topic. Currently, the RippleNet model has become a relatively effective knowledge graph-based recommendation model. However, the RippleNet model only uses the information between the user to be predicted and the item, and does not take into account the information contained in the neighbouring entities around the item to be predicted. To solve this problem, this paper proposes a new model based on RippleNet, Neig-DoubleRipple, which augments the original model by adding a neighbourhood attention mechanism and a neighbourhood entity aggregation method similar to the KGAT model. The experimental results show that the new model achieves 92.52% in auc and 85.11% in acc, indicating that the method proposed in this paper can significantly improve the performance of the RippleNet model.

In multivariate time series (MTS), missing data will lead to inaccurate predictions. In this paper, we propose a multivariate time series forecasting model based on bidirectional gated recurrent neural networks (RNN) combining self-attention with residual connections, abbreviated as BGRRSA. At the same time, the model also uses univariate time series prediction to extract historical information from the target series. Finally, the outputs of the multi-level self-attention mechanism and univariate time series prediction are weighted and summed to obtain the model's final prediction. Experiments were conducted to compare and analyze different missing values imputation methods and different time series prediction methods on two datasets, verifying the effectiveness and accuracy of the model.

Remote sensing data is a complex transformation process from data to information, from acquisition to application. These data come from a variety of sources and have a complex structure. It is extremely difficult for different researchers to access each other, which seriously affects the efficiency of scientific research. How to use artificial intelligence technology to manage these scattered files in a systematic multi-level classification is the key to remote sensing data management. In this paper, we propose a specially optimized file classification method based on deep learning technology, which aims to classify the random heterogeneous data generated by satellite remote sensing and expeditions in the field of fisheries expertise by means of artificial intelligence. The main contributions of this paper are threefold: (1) The binary data itself and its summary information in remote sensing data are input as different modalities for classification. (2) The BiLSTM model and the CNN model are improved for remote sensing data classification scenarios based on the multi-headed attention mechanism. (3) Feature extraction is performed based on the improved model and compared with the model of traditional classification methods.(3) The experimental results show that the improved file classification method has higher accuracy than the traditional machine learning classification methods in classifying fisheries remote sensing data files.

Smoky vehicle detection plays an important role in controlling urban pollution and protecting the ecological environment. Due to the complex shape transformation of smoke and the influence of environment, lighting and other factors, recognizing smoke from images is still a major challenge, and there have been many corresponding studies in recent years. Based on the above problems, we collected 2361 pictures of smoky cars as a smoke detection data set using the on-board camera. In addition, we used Yolov5 as the basic model and proposed an improved Yolov5 algorithm by preprocessing and expanding the data. The experimental results show that compared with the initial model, the smoke detection map of the improved algorithm is improved by 1.7%.

An optimized PCB welding sequence is important to improve both the welding rate and the safety of the robot. Both problems are essentially combinatorial optimization problems, and both are NP problems. In this paper, we use DRL for solving the end trajectory smoothing when the tin welding robot performs path planning on PCBs degree and distance problems for path planning on PCBs. In our approach, we use the concept of combinatorial optimization reward function to redefine the loss function of the PN. The better performance of using the proposed method compared to the manual heuristic algorithm is demonstrated through numerical experiments, in real PCBs scenario and in a simulation environment.

3D reconstruction from single hand-drawn sketches can be considered as a task of single-view reconstruction, which faces great challenge of lifting the dimension of the geometric representation of objects. Most of the reconstruction networks are based on deep learning technology with supervised training which are used to suffer from dataset labeling, while selfsupervised sketch-based 3D reconstruction remains challenging. In this paper, we propose a self-supervised 3D reconstruction network for hand-drawn sketch (IASSReNet), which introduces image information as an auxiliary to address the ambiguity and sparsity of sketch. In order to obtain image information, an image generator is firstly designed to provide augmented information for the reconstruction through a sketch feature enhancement module. To integrate the information from sketch and image, we use a spatially corresponding feature transfer module to fuse their feature. Finally, silhouettes are obtained from the predicted 3D mesh, and similarity constraints are applied to the sketch contour to perform self-supervised training on the network. Experimental results on multiple datasets show that our method outperforms other unsupervised methods and is competitive with some supervised methods.

Traditional relationship extraction requires manually labeling a small number of relationships and corresponding entities in a specific field as the training set. However, when switching to a new scene, it is necessary to redefine the rules and relabel the data. This will consume a lot of resources. With the continuous change of data types in the aviation field, it is necessary to design a general entity-relationship extraction model. For the research of open relation extraction in aviation field, we propose an open knowledge element extraction model based on Roberta + head-tail recognizer called RHTR. We constructed and made public a knowledge meta-corpus in the field of aviation. The model inputs the source text into Roberta to obtain the corresponding word embedding. The embedding is fed to the head-tail recognizer to do twice five classifications to get the head and tail sequence. Matching to the position information in the sentence based on the head and tail sequence, so as to get the corresponding label information. Then the label information is matched using the proximity principle to obtain the knowledge elements of the target. RHTR achieves 85.0% precision and 80.6% recall on the aviation field corpus. This fully verifies the superiority of the model. RHTR can obtain more kinds of knowledge elements while ensuring the accuracy of knowledge elements.

The current conventional fusion imaging system mainly relies on the fusion of light scattering information to obtain imaging results, which leads to poor fusion imaging results due to the lack of correction and denoising of images. In this regard, the design of an ISAR fusion imaging system based on an orthogonal matching tracking algorithm is proposed. A long-wave infrared detector is used as the main acquisition design for infrared images, and the hardware structure of the system is designed. The non-uniformity of the IR image is corrected by combining the single-point correction algorithm, and the image is filtered and de-noised, and finally the image pixels are fused by the pixel weighting method. In the experiments, the proposed fusion imaging system is verified for its imaging performance. The experimental results show that the proposed fusion imaging system performs image fusion with low imaging ranging errors and has a more satisfactory fusion imaging performance.

Efficient deployment of nodes inside the Wireless Sensor Network contributes to achieving maximum network coverage and minimum energy consumption. Aiming at the problems of low sensor node coverage and high redundancy of Sink nodes in the convergence layer in wireless sensor networks, we propose the WSN coverage optimization method based on the improved cuckoo search algorithm. First, the qubit Bloch spherical coordinates are introduced in the population initialization process to maintain better population diversity and wider coverage space. Secondly, in the Levy flight optimization stage of the cuckoo search algorithm, the residual operation inside the candidate solution matrix is used to generate long-distance dependency information of sensor nodes and new candidate solutions. Finally, a greedy evaluation strategy of dimension-by-dimension update is added to improve the random walk selection process to avoid the interference of node information between the same dimensions, so as to improve the iteration speed and accuracy of cuckoo search algorithm. Here, we use Matlab to build the coding environment and simulate the sensor nodes. The simulation results show that, compared with the leading methods in the current field, the improved cuckoo search algorithm can achieve a higher coverage ratio, and the redundancy ratio of Sink nodes in the aggregation layer is lower. Our proposed ICS model has better robustness and higher accuracy.

Due to their wide applications in imaging, display and optical communication, polarization singularities detection has become an inescapable task in these fields. However, because of the noise and turbulence imposed during the transmission of polarization information, traditional methodologies have difficulty in recognizing the disturbed polarization state effectively. The received polarization information may be de-correlated even the input polarization singularities are the same. Therefore, it remains a challenge to detect the polarization singularities in a fluctuated environment. In this work, we experimentally explore the potential of deep-learning in the classification and regression of polarization singularities. It has been demonstrated that this approach realizes the high accuracy of 98.89% in polarization singularity classification and gets low mean absolute error (MAE) ranges of ±0.021 in polarization singularity regression. The proposed approach proves the successful classification and regression of polarization singularities using deep learning methods.

With the continuous development of autonomous driving technology, more and more vehicles are equipped with different kinds of sensors for environmental sensing. Among these sensors, LiDAR and camera are the most visible, but there are great data structure differences between the point cloud data of LiDAR and RGB image data of camera, which bring some difficulties to the fusion. To address the above issues, this paper proposes a multimodal fusion framework CAD-Net for LiDAR-Camera fusion for 3D object detection. CAD-Net demonstrates the significant performance improvement of the proposed method by conducting comparative experiments and correlation analysis on KITTI public dataset.

In this paper, we propose an improved online self-organizing radial basis function neural network (IOS-RBF) modeling method which can dynamically add or merge hidden neuron while tuning the parameters. First, the initial center of the network is determined using the sample output clustering method, which is able to utilize the prior information contained in the output data. Second, a new neuron self-organization adjustment strategy is proposed, which is able to dynamically optimize the network structure according to the generalization ability of the network and the correlation between the nodes, and then, the network parameters are trained using the sliding window Levenberg-Marquardt (LM) algorithm to accelerate the network convergence speed. Finally, the effectiveness of IOS-RBF is demonstrated by two benchmark simulation experiments.

The application of deep learning-based methods in ship detection remains challenging due to the small size of the target of interest and the interference of complex sea surface background. Meanwhile, existing ship detection methods tend to experiment on a single image type, but rarely verify the applicability of their methods to multi-source sensor images. Therefore, this paper proposes a multi-scale deformable ship detector (MDSD) method based on the YOLOv3 framework for ship detection in optical and synthetic aperture radar (SAR) images. Since various information in the complex sea surface background can interfere with target detection, the detection algorithm designed in this paper includes a new multi-scale deformable attention module (MDAM), which can be added to the FPN part of the YOLOv3 model to enable the network to highlight target features while suppressing background features. The proposed MDAM generates attention maps from feature maps with multiscale deformable sensory fields, which can be better adapted to ship objects of various shapes and sizes. Compared with other models, MDSD has the highest detection accuracy on both optical image dataset (LEVIR) and SAR image dataset (SSDD). Also, MDSD is robust, which is validated on mixed datasets containing optical and SAR images.

This paper proposes a weighted hybrid forecasting model that combines the strengths of Prophet and SARIMAX models for predicting monthly electricity consumption. The case study focuses on northern China's city B, utilizing monthly electricity consumption and temperature data from 2010 to 2019. The weighted hybrid model achieves superior performance compared to individual Prophet and SARIMAX models, with Mean Absolute Percentage Error (MAPE) reduced by 22.8% and 19.0% respectively. The weighted hybrid model effectively integrates the advantages of both forecasting methods and significantly improves prediction accuracy.

The insulated gate bipolar transistor (IGBT) has the process of asymptotic and nonlinear degradation. It is of great significance to analyse its aging process and predict its remaining useful life for the safe and reliable operation of the power system. Therefore, this research introduces a method based on the Empirical Wavelet Transform combined with the Long Short Term Memory Network (EWT-LSTM) method to predict the remaining useful life of IGBT devices, to select the collector-emitter off instantaneous peak voltage as the characteristic signal of IGBT device aging. Firstly, the EWT method is used to decompose the aging signal. Then, in response to the problem of manually setting the number of segmentation layers in the EWT method, the adaptive spectrum segmentation method is used to adaptively determine the number of segmentation layers based on the signal to avoid interference caused by human participation. Finally, the LSTM time series model is used to predict each decomposed signal, and each prediction result is reconstructed to obtain the final prediction result. The results show that EWT-LSTM does not require noise filtering compared to other methods and has better prediction results, enabling better completion of IGBT life prediction work.

The widespread transmission of infectious diseases is a significant threat to human safety and development, and predicting epidemic trends is a pressing concern for scholars. Data assimilation methods can improve prediction accuracy by incorporating real observation data into system models. However, current data assimilation in epidemic models is limited to macro models, which fail to reflect individual heterogeneity realistically. Conversely, micro models have a high-dimensional parameter space, and the computational complexity of using data assimilation to adjust the state of each individual is high. To address this issue, this paper proposes a new data assimilation framework based on a macro-micro hierarchical simulation model called MaMiH. The MaMiH framework integrates the SEIR model as the macro model, reducing the computational complexity of the data assimilation process, and a complex network model as the micro model to simulate the infection status of each node within the network, thus reflecting individual heterogeneity. This paper verifies the effectiveness of the framework using the early stages of the COVID-19 pandemic as a research background.

In order to improve the accuracy of power dispatch professional language understanding, the professional language understanding method of power dispatching based on multi-model fusion is proposed. First, the dispatch professional language is represented as a low-dimensional feature vector based on the pre-trained word vector model. Then the mapping relationship between scheduling professional language and scheduling intention is trained based on text convolutional neural network (TextCNN). The relation relationship between professional language slot feature and information labels are trained based on the bidirectional long-term short-term memory network-conditional random field (BiLSTM-CRF), the dispatch professional language understanding is realized by the joint multi-model recognition results. Finally, through the verification of power dispatch professional language of a control center, compared with other methods, the proposed professional language understanding method has higher accuracy.

Highly accurate runoff forecasting is essential for the efficient use of water resources. Considering the non-linearity and randomness of runoff sequences, a selective ensemble forecasting method combined with the PSO algorithm is proposed. Firstly, sub-learners are selected for homogeneous ensemble and the parameters of each sub-learner are rate-determined on the training set. Then, the PSO algorithm is used to assign weights based on the performance of the sub-learners on the validation set to obtain the selective ensemble model. Finally, the selective ensemble method was validated on the test set. Experiments were performed using runoff data from the WuLong station in the Yangtze River basin, and the results show that the selective ensemble method can provide more accurate forecast results than homogeneous ensemble with the same average weights of the learners.